The research focused on two forecasting tasks. The first predicted hourly PM2.5 concentrations 6, 12 and 24 hours ahead, using lagged PM2.5 measurements, data from neighbouring stations, current and near-term temperature information, and other environmental variables. The second predicted 24-hour average PM2.5 concentrations 24, 48 and 72 hours ahead. Across the development period, the best short-term model at Station 1 was XGBoost, which retained an R² of 0.80 at 12 hours and 0.79 at 24 hours, while its RMSE rose from 6.59 to 7.09 and 7.48 micrograms per cubic metre. AdaBoost and CatBoost were also competitive, and all boosting approaches generally outperformed linear regression, the decision tree and the neural network in the short-term task.

Validation on the separate 2024 records produced a more mixed picture. At Station 1, AdaBoost had an R² of 0.81 and RMSE of 6.87 micrograms per cubic metre for the six-hour forecast, while Random Forest reached an R² of 0.80 and RMSE of 6.95 micrograms per cubic metre. At Station 2, Random Forest, CatBoost and AdaBoost remained stronger than the simpler models, although their R² values ranged from 0.50 to 0.66 for the 24-hour horizon. Station 3 showed R² values of about 0.64 to 0.67 and RMSE values of 8.35 to 8.69 micrograms per cubic metre for six-hour predictions with Random Forest and CatBoost, falling to roughly 0.29 to 0.57 at 24 hours depending on the algorithm. These results show that forecast skill depended on location, horizon and model rather than following a single ranking.

Daily averaging changed the pattern. For forecasts of 24-hour average PM2.5, Random Forest led Station 1 during development, with an R² of 0.72 and RMSE of 7.14 micrograms per cubic metres, and also performed strongly at Station 2, with an R² of 0.74 and RMSE of 7.17 micrograms per cubic metres. At Station 3, however, linear regression and the neural network each reached an R² of 0.82 and RMSE of 4.41 micrograms per cubic metres for both the 24- and 48-hour tasks. In the independent 2024 validation, linear regression at Station 1 achieved an R² of 0.85, RMSE of 5.45 micrograms per cubic metres and MAE of 3.66 micrograms per cubic metres for the 24-hour forecast. It remained at an R² of 0.84 for 48 hours before declining to 0.70 at 72 hours.

The same general result appeared at the other aggregated stations during validation. At Station 2, linear regression produced R² values of 0.81 and 0.80 for the 24- and 48-hour averaged forecasts, respectively, and 0.66 for 72 hours. At Station 3, linear regression and the neural network achieved R² values of 0.81 at 24 and 48 hours and 0.67 at 72 hours. Across other algorithms, the 72-hour R² values ranged from 0.45 to 0.65. The study therefore found that temporal averaging smoothed short-term fluctuations and reduced the performance gap between complex ensemble methods and simpler models, although accuracy still declined at the longest horizon.

Data preparation shaped the scope of the findings. Eight air-quality monitoring stations were consolidated into three geographically defined groups, with hourly pollutant values averaged within each group. The authors said this supported city-scale analysis but also masked local pollution hotspots and microscale spatial differences. The Tabriz East meteorological station was excluded because of data gaps, limited variables and high error rates in automated recordings; the Tabriz Airport station supplied the meteorological input. Air-quality records were combined with three-hourly and daily weather observations, monthly natural-gas consumption and 16-day Normalized Difference Vegetation Index composites derived from Landsat 8 and 9 imagery. Missing pollutant values were imputed with XGBoost before lagged features were created and before the 2013–2023 and 2024 periods were separated.

The historical record showed pronounced seasonality in Tabriz. Mean daily PM2.5 concentrations across 2013–2023 were 23.28 micrograms per cubic metres at Station 1, 21.19 at Station 2 and 18.26 at Station 3. The authors associated higher cold-season concentrations with residential heating, low wind speeds, limited atmospheric mixing and temperature inversions, conditions favored by the city’s mountainous basin setting. The model interpretability analysis identified lagged PM2.5 and horizontal visibility as leading predictors for 6- and 12-hour forecasts. Temperature-related variables became more influential at 12 to 24 hours, while pressure, solar radiation and sunshine duration showed horizon-dependent contributions.

Natural-gas consumption provided an additional local predictor, especially at Station 3 and at longer horizons. SHAP-based analysis showed that higher gas use was generally associated with higher predicted PM2.5, a pattern the authors interpreted as consistent with residential-heating emissions. Visibility had a strong effect at Station 1, with lower visibility associated with higher predictions, while solar radiation, sunshine duration and vegetation indices became more visible in some longer-horizon models. The study did not establish that any single variable caused pollution episodes, and the feature analyses were performed on the best-performing models rather than on every algorithm and station combination. Interpretability therefore describes how the fitted models used their inputs, not a complete causal account of Tabriz’s pollution sources.

Several limitations constrain how the results should be interpreted. The aggregated stations cannot represent neighborhood-scale pollution, and the study relied on ground-based observations without satellite aerosol optical depth or boundary-layer height. Hourly gas values were created by dividing monthly consumption by 30 and copying that value across each day, which suppresses within-day variation, while other lower-frequency meteorological and vegetation measurements were also replicated rather than interpolated. The authors acknowledged a possible information-leakage risk because imputation was performed on the assembled dataset before separating the development and 2024 test periods, even though the predictive models themselves were trained only on 2013–2023 data. Conventional random five-fold cross-validation also did not preserve temporal order and may have produced optimistic estimates, and no prediction intervals were reported.

Traffic-density data and fossil-fuel consumption for gasoline, diesel and heavy fuel oil were unavailable, preventing the models from representing those sources directly. The results are therefore best understood as a city-scale forecasting study based on the available monitoring network rather than a complete urban emissions inventory or a locally validated early-warning system. The tested algorithms included Random Forest, Gradient Boosting, XGBoost, CatBoost, AdaBoost, a decision tree, linear regression and a neural network, with performance evaluated through R², RMSE and MAE. The central result was practical rather than universal: boosting ensembles were consistently among the strongest choices for short-range forecasts, while averaging over longer periods allowed simpler models to match or exceed them at some stations. Any future operational use would require temporally ordered validation, independent fitting of the imputation procedure, uncertainty intervals, updated inputs and evaluation at a finer spatial scale.

Subject of Research: Machine-learning forecasts of short- and multi-day PM2.5 concentrations in Tabriz, Iran

Article Title: Machine learning approaches for short- and multi-day PM 2.5 prediction in Tabriz, Iran

Article References: Alinezhad, N., Samad-Soltani, T., Sabeti, Z., & Gholampour, A. (2026). Machine learning approaches for short- and multi-day PM2.5 prediction in Tabriz, Iran. Heliyon, 12(16), Article e45597. https://doi.org/10.1016/j.heliyon.2026.e45597

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45597

Keywords: PM2.5 forecasting, Tabriz, machine learning, XGBoost, air pollution, meteorology, time-series models, environmental monitoring, Machine, learning, approaches, short-